Yes. A DGX Station GB300 has one GPU, one 1,600 W-class power supply, one liquid-cooling loop and, on several configurations, a data volume striped across two drives with no redundancy. If a whole AI team depends on it and nothing else, any of those can stop the team's work. The fix is not a second Station by default. It is a short list of controls: a backup copy of everything you cannot download again, a power plan, a hardware support term you have read, and a fallback for urgent work while the unit is being repaired. The HP ZGX Fury G1n, MSI XpertStation WS300, ASUS ExpertCenter Pro ET900N G3, GIGABYTE W775-V10-L01 and Supermicro ARS-511GD-NB-LCC are price on request at SourceIT, and all five share NVIDIA's single-superchip design.
Where can a DGX Station fail, and what protects against it?
| Component | Built-in protection | What you still need |
|---|---|---|
| OS drives | Two Gen5 NVMe drives in software RAID 1 (NVIDIA reference layout) | An OS reinstall plan and configuration notes; RAID 1 does not protect against deletion, corruption or ransomware |
| Data drives | RAID 0 on the HP and Supermicro layouts (no redundancy); empty slots on MSI and base ASUS | Copies of datasets, fine-tuned models and checkpoints on separate storage |
| GPU and CPU | One GB300 superchip; no second GPU to fail over to | A fallback for urgent work and a hardware support term |
| Memory | 496GB LPDDR5X on four SOCAMM modules under cold plates | OEM service; MSI's manual says opening that area voids the warranty |
| Power supply | One 1,600 W-class unit; HP's listing calls for a dedicated 200–240 V circuit | A UPS and a suitable circuit |
| Cooling | Liquid cooling is mandatory; only GIGABYTE lists leak detection with automatic shutdown | Temperature and fault monitoring, and a clean, ventilated location |
| Hardware support | HP and Supermicro listings state 3-year onsite terms; MSI, ASUS and GIGABYTE confirm terms at quotation | A written response time and an escalation contact |
Sources: NVIDIA DGX Station Development Guide (30 Oct 2026), ServeTheHome's MSI WS300 review (23 Oct 2026), and SourceIT's five DGX Station listings (4 Nov 2026). We have not tested these systems ourselves.
Isn't RAID enough?
No. RAID keeps a system running when one drive fails. It does not keep a second copy of your data. If someone deletes a folder, a script corrupts a checkpoint or ransomware encrypts the disks, every mirrored drive gets the same change at the same moment. On the HP and Supermicro layouts the data drives are striped (RAID 0), so a single data-drive failure loses the whole volume. That is a sensible design for speed, but it makes off-box copies essential. Our DGX Station storage article shows the usable space on each OEM's layout.
What should the backup plan cover?
Singapore's CSA Cyber Essentials mark (2nd edition, March 2025) gives a practical baseline. It asks organisations to:
- back up business-critical data and systems, at a frequency "aligned to business requirements";
- store backups "separately and isolated from the operating environment";
- protect backups, at minimum with a passphrase;
- test restores "at least bi-annually".
On a DGX Station, sort the data into three groups. Public model weights and containers can be downloaded again, so a list of versions is enough. Datasets you prepared, fine-tuned models, prompts, evaluation results and checkpoints cannot, so back them up off the box. System configuration, such as user accounts, MIG layouts, service files and firewall rules, should be documented or kept in version control so a rebuilt system can be restored quickly.
Size matters here. A common rule of thumb from the ZeRO paper puts a full mixed-precision training state at about 16 bytes per parameter, which is about 1.12TB for a 70-billion-parameter model. Backing that up over a 1GbE office link is slow, so connect backup storage at 10GbE. Every DGX Station has a 10GbE RJ45 port.
If losing the Station for a week would stop your AI team, the Station is critical infrastructure and should be backed up like it.
What happens while the unit is being repaired?
Decide this before it happens. There is no second GPU inside the Station to fail over to, so the options are:
- A smaller fallback system. A GB10 system such as the NVIDIA DGX Spark 128GB/4TB (S$9,300) or HP ZGX Nano G1n 128GB/4TB (S$11,499, both listed 4 Nov 2026) can keep smaller models and development work going. Its 128GB of memory cannot hold the largest models the Station runs, so treat it as a reduced service.
- A second Station. NVIDIA documents linking two Stations back to back over their 400Gb/s ports. The pair can share large jobs, and either unit can carry on alone if the other fails. This suits teams whose revenue depends on the service.
- Temporary cloud GPUs. These are useful for short outages, provided your data rules allow it. Our article on where company data goes when using AI covers that question.
Check the warranty wording too. SourceIT's HP ZGX Fury listing states 3 years of parts, labour and onsite repair. The Supermicro listing states a 3-year next-business-day onsite term. MSI, ASUS and GIGABYTE terms are confirmed in writing at quotation. Our HP vs MSI warranty and support comparison goes deeper on that.
How do you protect the power and cooling?
NVIDIA rates the Station at 1,600 W. GIGABYTE reports a typical 1,000–1,200 W under load on its version. A power cut during a long training run can lose hours of work unless checkpoints are saved regularly, and an unclean shutdown is unkind to any system. Put the Station on an online (double-conversion) UPS sized for the full 1,600 W rating with headroom. In stock at SourceIT are the APC Smart-UPS SRT 3000VA 230V tower at S$4,165 and the APC Smart-UPS On-Line 3kVA 2U rackmount at S$4,480 (both listed 4 Nov 2026). Ask us to confirm the watt rating and runtime for your load. Our articles on protecting an expensive workstation with a UPS and UPS batteries that fail quietly explain why the UPS itself needs testing.
For cooling, monitor GPU temperature and errors through DCGM, which is pre-installed, and the BMC's out-of-band telemetry. Keep the intake clear. The office installation checklist covers placement, circuits and airflow.
What about ransomware and human error?
These are the failures RAID cannot fix. Keep at least one backup copy that the Station's own accounts cannot delete, because CSA asks for backups to be isolated from the operating environment. Use named accounts and limit sudo, so one mistake or one stolen password cannot wipe everything. Our articles on keeping the business running during ransomware and whether the Station should be on the internet cover the security side. Who manages the Station covers ownership.
Buying in Singapore
SourceIT supplies the HP ZGX Fury G1n, MSI XpertStation WS300, ASUS ExpertCenter Pro ET900N G3, GIGABYTE W775-V10-L01 and Supermicro ARS-511GD-NB-LCC on a price-on-request basis. Supporting kit in stock includes the APC Smart-UPS SRT 3000VA (S$4,165), APC Smart-UPS On-Line 3kVA rackmount (S$4,480), Synology RS822RP+ NAS (S$2,086.80), NVIDIA DGX Spark (S$9,300) and HP ZGX Nano G1n (S$11,499), all listed 4 Nov 2026. Browse the NVIDIA DGX Spark and DGX Station collection and UPS and power protection. For a quotation, use Request for Quotation on the product page, call (65) 6978 3502, email Sales@sourceit.com.sg or use our contact page.
Frequently asked questions
Does a DGX Station have redundant parts? Only the OS drives, which NVIDIA's reference layout mirrors in RAID 1. There is one GB300 superchip, one 1,600 W-class power supply and one liquid-cooling loop.
Is RAID 1 a backup? No. It protects against a single drive failure, but deletions, corruption and ransomware affect both mirrored drives at once. Keep separate, isolated backups.
How often should backups be tested? Singapore's CSA Cyber Essentials mark asks for restores to be tested at least twice a year.
What can we use while the Station is being repaired? A GB10 system such as the DGX Spark (S$9,300) or HP ZGX Nano (S$11,499) can run smaller models and development work. A second Station or temporary cloud GPUs cover larger jobs.
Does a DGX Station need a UPS? It is strongly advisable. It is a 1,600 W system that often runs long jobs, so an online UPS sized for that load protects both the hardware and work in progress.
SourceIT Approved
Researched and verified by the SourceIT team for Singapore buyers.
